Gtm Machine

GTM Machine

Turn a startup idea into evidence, and evidence into booked calls. Five stages, each with a hard gate. The north-star metric is booked calls per idea tested — not reports generated, not leads sourced, not emails sent.

Honest status of this skill

The process encoded here passed its own evals (11/11 on a golden set of ideas, live-validated Apollo ICPs, a verified end-to-end send-and-classify loop). What it has NOT yet done is book a call — the full outbound loop was proven mechanically, not commercially. Treat the reply-rate bars below as starting hypotheses, and tell the founder this when setting expectations. A validation tool that oversells itself commits the exact sin it exists to catch.

The one rule that outranks everything

Never state anything about a prospect's business that you have not verified. A cold email that says "your chat widget does X" to a hotel with no chat widget is dead on arrival — worse than generic, because being wrong proves you never looked. Every claim about a recipient needs a scan record with evidence and a timestamp, or it does not ship. This rule was learned the expensive way; enforce it programmatically (see scripts/s4_write/lint.py, the CLAIM_FIELDS gate), never by good intentions.

The five stages

Work through them in order. Each stage's full rubric lives in references/ — read the relevant file when you reach that stage, not before.

S1 — Validate → references/s1-validate.md

Score the idea on demand, AI-platform kill-risk, and GTM viability (1–5 each), verdict pursue/park/kill. Every score cites gathered evidence by id — a score without a trace is invalid output. Honest by construction: it must be willing to kill the founder's favorite idea, disagree openly, and score ideas resembling its own creator's without flattery. Output: a shareable teardown report. Runnable module: scripts/s1_validate/ + golden-set eval in scripts/eval/.

S2 — Hypothesize → references/s2-hypothesize.md

Convert a pursue verdict into 2–3 personas and 1–3 prioritized, falsifiable outbound hypotheses, each with an ICP emitted directly as Apollo search parameters (validated live: 100–2,000 people or the filter is wrong). If incumbents exist — and they almost always do — at least one hypothesis must test displacement, not greenfield. Module: scripts/s2_hypothesize/.

S3 — Source → references/s3-source.md

Execute the ICP against Apollo: search free, enrich only post-filter survivors (credits are the budget), score fit 1–5 with a reason, dedupe by sending domain (not company name), and site-scan each lead to verify the claims the copy will make. Leads failing the scan gate are blocked, not sent generic. Module: scripts/s3_source/.

S4 — Write → references/s4-write.md

2–3 first-touch variants testing different mechanisms (observation-led, number-led, question-led), ≤90 words rendered, one CTA with a booking link, an opt-out line, zero spam triggers. Lint every rendered (email × lead) pair — template-level linting is insufficient because merge fields expand. Module: scripts/s4_write/.

S5 — Send & Learn → references/s5-send.md

3-touch sequence (day 0, +3, breakup at +10), state machine over reply classes, intent and sentiment classified separately, last-touch attribution. Nothing sends without a fresh, batch-scoped human approval — approving touch 1 grants no authority over touch 2. The approval binds to the exact lead set and a hash of the rendered copy; any change kills it. Modules: scripts/s5_send/ (gate.py is the enforcement, classify.py the reply classifier).

Hard gates (enforced in code, not vibes)

1. Approval per batch. The scheduler prepares and notifies; it has no send path. scripts/s5_send/gate.py — refuses unapproved, mutated, or stale (>24h) batches.
2. Verified claims only. No property_observation without scan_evidence. The scanner fails closed: if it cannot render the page (JS execution required — chat widgets are runtime-injected), it blocks the lead rather than reporting a false "no widget".
3. negative ≠ opt_out. "No, thanks" suppresses one hypothesis; "remove me" suppresses globally and forever. Over-suppressing polite refusals silently destroys future addressable market; under-suppressing removals is a legal breach. Deterministic regex catches opt-outs/bounces before any LLM call.
4. Caps belong to the mailbox, not the campaign. ≤30 sends/day across ALL campaigns; one live thread per sending domain; bounce >3% per rolling 50 auto-pauses; a kill-switch cell halts everything.
5. Compliance by geography. UK PECR / IE S.I. 336: corporate subscribers may be cold-emailed with identification + opt-out; sole traders need consent. Registry-check any lead below ~50 employees (scripts/compliance/soletrader_check.py — fails closed). Not legal advice; say so.

System of record

Everything the machine does must be inspectable by the founder without asking. Default: one Drive folder per idea (teardown, hypothesis pack, campaign copy) plus sheets for ideas / hypotheses / leads / campaigns / sends / replies — schemas in references/operations.md. Every verdict, score, and send carries its evidence and timestamps. If the founder can't audit it, don't do it.

Connections & setup → references/setup.md

On first use — and at every stage boundary — check what's connected and hand-hold the founder through what's missing: Apollo (plan tier, API key, mailbox link, signature toggle), Google Drive (and its create-only limitation), email read vs send paths, Firecrawl for the scan gate, and the script env keys. references/setup.md has the per-connector walkthroughs, what each stage requires vs what degrades gracefully, and a first-run verification checklist (probe everything live; never trust "connected" status). Two principles: S1 needs nothing — never block idea validation on outbound tooling; and state costs out loud before spending anything (Apollo credits, API tiers).

Where to start with a founder

1. Ask for the idea (or offer to suggest ideas in a domain they know). Check connections per references/setup.md — but start S1 regardless.
2. Run S1. Show the teardown, including the uncomfortable parts.
3. If pursue: S2 hypotheses, and push the founder to do the one thing the machine cannot — talk to 4–5 real buyers. Interview evidence outranks everything scraped.
4. S3–S5 only with their budget consent (Apollo credits) and per-batch send approval.
5. Record outcomes. The accumulated angle × ICP × conversion data is the compounding asset — instrument it from campaign one.